Pedestrian Detection Using Optimized YOLOv3 in UAV Scenario
Changhao Piao, Xianhao Wang, Mingjie Liu · 2019
Pedestrian detection is of great importance in computer vision task. It is a hard work since the performance is affected by environmental factors such as illumination, occlusion, and background clutters, especially in UAV scenario where the pedestrian is always with a small scale. In this paper, we propose adaptive random multi-scale variation method to detect small object in UAV scenario. First, the special pedestrian dataset used for UAV is collected to train the proposed method. Then, we improve YOLOv3, which is a classical one-stage object detection method, by data sample classification (normal, far, game scenario), adaptive random multi-scale variation, and k-means clustering to make the model robust to scale variation. Compared with the original YOLOv3 pedestrian detection result, the proposed model has a high detection precision (90.91%) and accuracy (80.59%).